2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
5 * The allocator synchronizes using per slab locks and only
6 * uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter <clameter@sgi.com>
12 #include <linux/module.h>
13 #include <linux/bit_spinlock.h>
14 #include <linux/interrupt.h>
15 #include <linux/bitops.h>
16 #include <linux/slab.h>
17 #include <linux/seq_file.h>
18 #include <linux/cpu.h>
19 #include <linux/cpuset.h>
20 #include <linux/mempolicy.h>
21 #include <linux/ctype.h>
22 #include <linux/kallsyms.h>
23 #include <linux/memory.h>
30 * The slab_lock protects operations on the object of a particular
31 * slab and its metadata in the page struct. If the slab lock
32 * has been taken then no allocations nor frees can be performed
33 * on the objects in the slab nor can the slab be added or removed
34 * from the partial or full lists since this would mean modifying
35 * the page_struct of the slab.
37 * The list_lock protects the partial and full list on each node and
38 * the partial slab counter. If taken then no new slabs may be added or
39 * removed from the lists nor make the number of partial slabs be modified.
40 * (Note that the total number of slabs is an atomic value that may be
41 * modified without taking the list lock).
43 * The list_lock is a centralized lock and thus we avoid taking it as
44 * much as possible. As long as SLUB does not have to handle partial
45 * slabs, operations can continue without any centralized lock. F.e.
46 * allocating a long series of objects that fill up slabs does not require
49 * The lock order is sometimes inverted when we are trying to get a slab
50 * off a list. We take the list_lock and then look for a page on the list
51 * to use. While we do that objects in the slabs may be freed. We can
52 * only operate on the slab if we have also taken the slab_lock. So we use
53 * a slab_trylock() on the slab. If trylock was successful then no frees
54 * can occur anymore and we can use the slab for allocations etc. If the
55 * slab_trylock() does not succeed then frees are in progress in the slab and
56 * we must stay away from it for a while since we may cause a bouncing
57 * cacheline if we try to acquire the lock. So go onto the next slab.
58 * If all pages are busy then we may allocate a new slab instead of reusing
59 * a partial slab. A new slab has noone operating on it and thus there is
60 * no danger of cacheline contention.
62 * Interrupts are disabled during allocation and deallocation in order to
63 * make the slab allocator safe to use in the context of an irq. In addition
64 * interrupts are disabled to ensure that the processor does not change
65 * while handling per_cpu slabs, due to kernel preemption.
67 * SLUB assigns one slab for allocation to each processor.
68 * Allocations only occur from these slabs called cpu slabs.
70 * Slabs with free elements are kept on a partial list and during regular
71 * operations no list for full slabs is used. If an object in a full slab is
72 * freed then the slab will show up again on the partial lists.
73 * We track full slabs for debugging purposes though because otherwise we
74 * cannot scan all objects.
76 * Slabs are freed when they become empty. Teardown and setup is
77 * minimal so we rely on the page allocators per cpu caches for
78 * fast frees and allocs.
80 * Overloading of page flags that are otherwise used for LRU management.
82 * PageActive The slab is frozen and exempt from list processing.
83 * This means that the slab is dedicated to a purpose
84 * such as satisfying allocations for a specific
85 * processor. Objects may be freed in the slab while
86 * it is frozen but slab_free will then skip the usual
87 * list operations. It is up to the processor holding
88 * the slab to integrate the slab into the slab lists
89 * when the slab is no longer needed.
91 * One use of this flag is to mark slabs that are
92 * used for allocations. Then such a slab becomes a cpu
93 * slab. The cpu slab may be equipped with an additional
94 * freelist that allows lockless access to
95 * free objects in addition to the regular freelist
96 * that requires the slab lock.
98 * PageError Slab requires special handling due to debug
99 * options set. This moves slab handling out of
100 * the fast path and disables lockless freelists.
103 #define FROZEN (1 << PG_active)
105 #ifdef CONFIG_SLUB_DEBUG
106 #define SLABDEBUG (1 << PG_error)
111 static inline int SlabFrozen(struct page *page)
113 return page->flags & FROZEN;
116 static inline void SetSlabFrozen(struct page *page)
118 page->flags |= FROZEN;
121 static inline void ClearSlabFrozen(struct page *page)
123 page->flags &= ~FROZEN;
126 static inline int SlabDebug(struct page *page)
128 return page->flags & SLABDEBUG;
131 static inline void SetSlabDebug(struct page *page)
133 page->flags |= SLABDEBUG;
136 static inline void ClearSlabDebug(struct page *page)
138 page->flags &= ~SLABDEBUG;
142 * Issues still to be resolved:
144 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
146 * - Variable sizing of the per node arrays
149 /* Enable to test recovery from slab corruption on boot */
150 #undef SLUB_RESILIENCY_TEST
155 * Small page size. Make sure that we do not fragment memory
157 #define DEFAULT_MAX_ORDER 1
158 #define DEFAULT_MIN_OBJECTS 4
163 * Large page machines are customarily able to handle larger
166 #define DEFAULT_MAX_ORDER 2
167 #define DEFAULT_MIN_OBJECTS 8
172 * Mininum number of partial slabs. These will be left on the partial
173 * lists even if they are empty. kmem_cache_shrink may reclaim them.
175 #define MIN_PARTIAL 5
178 * Maximum number of desirable partial slabs.
179 * The existence of more partial slabs makes kmem_cache_shrink
180 * sort the partial list by the number of objects in the.
182 #define MAX_PARTIAL 10
184 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
185 SLAB_POISON | SLAB_STORE_USER)
188 * Set of flags that will prevent slab merging
190 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
191 SLAB_TRACE | SLAB_DESTROY_BY_RCU)
193 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
196 #ifndef ARCH_KMALLOC_MINALIGN
197 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
200 #ifndef ARCH_SLAB_MINALIGN
201 #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
204 /* Internal SLUB flags */
205 #define __OBJECT_POISON 0x80000000 /* Poison object */
206 #define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */
207 #define __KMALLOC_CACHE 0x20000000 /* objects freed using kfree */
208 #define __PAGE_ALLOC_FALLBACK 0x10000000 /* Allow fallback to page alloc */
210 /* Not all arches define cache_line_size */
211 #ifndef cache_line_size
212 #define cache_line_size() L1_CACHE_BYTES
215 static int kmem_size = sizeof(struct kmem_cache);
218 static struct notifier_block slab_notifier;
222 DOWN, /* No slab functionality available */
223 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
224 UP, /* Everything works but does not show up in sysfs */
228 /* A list of all slab caches on the system */
229 static DECLARE_RWSEM(slub_lock);
230 static LIST_HEAD(slab_caches);
233 * Tracking user of a slab.
236 void *addr; /* Called from address */
237 int cpu; /* Was running on cpu */
238 int pid; /* Pid context */
239 unsigned long when; /* When did the operation occur */
242 enum track_item { TRACK_ALLOC, TRACK_FREE };
244 #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
245 static int sysfs_slab_add(struct kmem_cache *);
246 static int sysfs_slab_alias(struct kmem_cache *, const char *);
247 static void sysfs_slab_remove(struct kmem_cache *);
250 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
251 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
253 static inline void sysfs_slab_remove(struct kmem_cache *s)
260 static inline void stat(struct kmem_cache_cpu *c, enum stat_item si)
262 #ifdef CONFIG_SLUB_STATS
267 /********************************************************************
268 * Core slab cache functions
269 *******************************************************************/
271 int slab_is_available(void)
273 return slab_state >= UP;
276 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
279 return s->node[node];
281 return &s->local_node;
285 static inline struct kmem_cache_cpu *get_cpu_slab(struct kmem_cache *s, int cpu)
288 return s->cpu_slab[cpu];
294 /* Verify that a pointer has an address that is valid within a slab page */
295 static inline int check_valid_pointer(struct kmem_cache *s,
296 struct page *page, const void *object)
303 base = page_address(page);
304 if (object < base || object >= base + page->objects * s->size ||
305 (object - base) % s->size) {
313 * Slow version of get and set free pointer.
315 * This version requires touching the cache lines of kmem_cache which
316 * we avoid to do in the fast alloc free paths. There we obtain the offset
317 * from the page struct.
319 static inline void *get_freepointer(struct kmem_cache *s, void *object)
321 return *(void **)(object + s->offset);
324 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
326 *(void **)(object + s->offset) = fp;
329 /* Loop over all objects in a slab */
330 #define for_each_object(__p, __s, __addr) \
331 for (__p = (__addr); __p < (__addr) + (__s)->objects * (__s)->size;\
335 #define for_each_free_object(__p, __s, __free) \
336 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
338 /* Determine object index from a given position */
339 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
341 return (p - addr) / s->size;
344 #ifdef CONFIG_SLUB_DEBUG
348 #ifdef CONFIG_SLUB_DEBUG_ON
349 static int slub_debug = DEBUG_DEFAULT_FLAGS;
351 static int slub_debug;
354 static char *slub_debug_slabs;
359 static void print_section(char *text, u8 *addr, unsigned int length)
367 for (i = 0; i < length; i++) {
369 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
372 printk(KERN_CONT " %02x", addr[i]);
374 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
376 printk(KERN_CONT " %s\n", ascii);
383 printk(KERN_CONT " ");
387 printk(KERN_CONT " %s\n", ascii);
391 static struct track *get_track(struct kmem_cache *s, void *object,
392 enum track_item alloc)
397 p = object + s->offset + sizeof(void *);
399 p = object + s->inuse;
404 static void set_track(struct kmem_cache *s, void *object,
405 enum track_item alloc, void *addr)
410 p = object + s->offset + sizeof(void *);
412 p = object + s->inuse;
417 p->cpu = smp_processor_id();
418 p->pid = current ? current->pid : -1;
421 memset(p, 0, sizeof(struct track));
424 static void init_tracking(struct kmem_cache *s, void *object)
426 if (!(s->flags & SLAB_STORE_USER))
429 set_track(s, object, TRACK_FREE, NULL);
430 set_track(s, object, TRACK_ALLOC, NULL);
433 static void print_track(const char *s, struct track *t)
438 printk(KERN_ERR "INFO: %s in ", s);
439 __print_symbol("%s", (unsigned long)t->addr);
440 printk(" age=%lu cpu=%u pid=%d\n", jiffies - t->when, t->cpu, t->pid);
443 static void print_tracking(struct kmem_cache *s, void *object)
445 if (!(s->flags & SLAB_STORE_USER))
448 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
449 print_track("Freed", get_track(s, object, TRACK_FREE));
452 static void print_page_info(struct page *page)
454 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
455 page, page->objects, page->inuse, page->freelist, page->flags);
459 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
465 vsnprintf(buf, sizeof(buf), fmt, args);
467 printk(KERN_ERR "========================================"
468 "=====================================\n");
469 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
470 printk(KERN_ERR "----------------------------------------"
471 "-------------------------------------\n\n");
474 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
480 vsnprintf(buf, sizeof(buf), fmt, args);
482 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
485 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
487 unsigned int off; /* Offset of last byte */
488 u8 *addr = page_address(page);
490 print_tracking(s, p);
492 print_page_info(page);
494 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
495 p, p - addr, get_freepointer(s, p));
498 print_section("Bytes b4", p - 16, 16);
500 print_section("Object", p, min(s->objsize, 128));
502 if (s->flags & SLAB_RED_ZONE)
503 print_section("Redzone", p + s->objsize,
504 s->inuse - s->objsize);
507 off = s->offset + sizeof(void *);
511 if (s->flags & SLAB_STORE_USER)
512 off += 2 * sizeof(struct track);
515 /* Beginning of the filler is the free pointer */
516 print_section("Padding", p + off, s->size - off);
521 static void object_err(struct kmem_cache *s, struct page *page,
522 u8 *object, char *reason)
524 slab_bug(s, "%s", reason);
525 print_trailer(s, page, object);
528 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
534 vsnprintf(buf, sizeof(buf), fmt, args);
536 slab_bug(s, "%s", buf);
537 print_page_info(page);
541 static void init_object(struct kmem_cache *s, void *object, int active)
545 if (s->flags & __OBJECT_POISON) {
546 memset(p, POISON_FREE, s->objsize - 1);
547 p[s->objsize - 1] = POISON_END;
550 if (s->flags & SLAB_RED_ZONE)
551 memset(p + s->objsize,
552 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
553 s->inuse - s->objsize);
556 static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
559 if (*start != (u8)value)
567 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
568 void *from, void *to)
570 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
571 memset(from, data, to - from);
574 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
575 u8 *object, char *what,
576 u8 *start, unsigned int value, unsigned int bytes)
581 fault = check_bytes(start, value, bytes);
586 while (end > fault && end[-1] == value)
589 slab_bug(s, "%s overwritten", what);
590 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
591 fault, end - 1, fault[0], value);
592 print_trailer(s, page, object);
594 restore_bytes(s, what, value, fault, end);
602 * Bytes of the object to be managed.
603 * If the freepointer may overlay the object then the free
604 * pointer is the first word of the object.
606 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
609 * object + s->objsize
610 * Padding to reach word boundary. This is also used for Redzoning.
611 * Padding is extended by another word if Redzoning is enabled and
614 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
615 * 0xcc (RED_ACTIVE) for objects in use.
618 * Meta data starts here.
620 * A. Free pointer (if we cannot overwrite object on free)
621 * B. Tracking data for SLAB_STORE_USER
622 * C. Padding to reach required alignment boundary or at mininum
623 * one word if debugging is on to be able to detect writes
624 * before the word boundary.
626 * Padding is done using 0x5a (POISON_INUSE)
629 * Nothing is used beyond s->size.
631 * If slabcaches are merged then the objsize and inuse boundaries are mostly
632 * ignored. And therefore no slab options that rely on these boundaries
633 * may be used with merged slabcaches.
636 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
638 unsigned long off = s->inuse; /* The end of info */
641 /* Freepointer is placed after the object. */
642 off += sizeof(void *);
644 if (s->flags & SLAB_STORE_USER)
645 /* We also have user information there */
646 off += 2 * sizeof(struct track);
651 return check_bytes_and_report(s, page, p, "Object padding",
652 p + off, POISON_INUSE, s->size - off);
655 /* Check the pad bytes at the end of a slab page */
656 static int slab_pad_check(struct kmem_cache *s, struct page *page)
664 if (!(s->flags & SLAB_POISON))
667 start = page_address(page);
668 length = (PAGE_SIZE << s->order);
669 end = start + length;
670 remainder = length % s->size;
674 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
677 while (end > fault && end[-1] == POISON_INUSE)
680 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
681 print_section("Padding", end - remainder, remainder);
683 restore_bytes(s, "slab padding", POISON_INUSE, start, end);
687 static int check_object(struct kmem_cache *s, struct page *page,
688 void *object, int active)
691 u8 *endobject = object + s->objsize;
693 if (s->flags & SLAB_RED_ZONE) {
695 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
697 if (!check_bytes_and_report(s, page, object, "Redzone",
698 endobject, red, s->inuse - s->objsize))
701 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
702 check_bytes_and_report(s, page, p, "Alignment padding",
703 endobject, POISON_INUSE, s->inuse - s->objsize);
707 if (s->flags & SLAB_POISON) {
708 if (!active && (s->flags & __OBJECT_POISON) &&
709 (!check_bytes_and_report(s, page, p, "Poison", p,
710 POISON_FREE, s->objsize - 1) ||
711 !check_bytes_and_report(s, page, p, "Poison",
712 p + s->objsize - 1, POISON_END, 1)))
715 * check_pad_bytes cleans up on its own.
717 check_pad_bytes(s, page, p);
720 if (!s->offset && active)
722 * Object and freepointer overlap. Cannot check
723 * freepointer while object is allocated.
727 /* Check free pointer validity */
728 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
729 object_err(s, page, p, "Freepointer corrupt");
731 * No choice but to zap it and thus loose the remainder
732 * of the free objects in this slab. May cause
733 * another error because the object count is now wrong.
735 set_freepointer(s, p, NULL);
741 static int check_slab(struct kmem_cache *s, struct page *page)
745 VM_BUG_ON(!irqs_disabled());
747 if (!PageSlab(page)) {
748 slab_err(s, page, "Not a valid slab page");
752 maxobj = (PAGE_SIZE << compound_order(page)) / s->size;
753 if (page->objects > maxobj) {
754 slab_err(s, page, "objects %u > max %u",
755 s->name, page->objects, maxobj);
758 if (page->inuse > page->objects) {
759 slab_err(s, page, "inuse %u > max %u",
760 s->name, page->inuse, page->objects);
763 /* Slab_pad_check fixes things up after itself */
764 slab_pad_check(s, page);
769 * Determine if a certain object on a page is on the freelist. Must hold the
770 * slab lock to guarantee that the chains are in a consistent state.
772 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
775 void *fp = page->freelist;
778 while (fp && nr <= page->objects) {
781 if (!check_valid_pointer(s, page, fp)) {
783 object_err(s, page, object,
784 "Freechain corrupt");
785 set_freepointer(s, object, NULL);
788 slab_err(s, page, "Freepointer corrupt");
789 page->freelist = NULL;
790 page->inuse = page->objects;
791 slab_fix(s, "Freelist cleared");
797 fp = get_freepointer(s, object);
801 if (page->inuse != page->objects - nr) {
802 slab_err(s, page, "Wrong object count. Counter is %d but "
803 "counted were %d", page->inuse, page->objects - nr);
804 page->inuse = page->objects - nr;
805 slab_fix(s, "Object count adjusted.");
807 return search == NULL;
810 static void trace(struct kmem_cache *s, struct page *page, void *object, int alloc)
812 if (s->flags & SLAB_TRACE) {
813 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
815 alloc ? "alloc" : "free",
820 print_section("Object", (void *)object, s->objsize);
827 * Tracking of fully allocated slabs for debugging purposes.
829 static void add_full(struct kmem_cache_node *n, struct page *page)
831 spin_lock(&n->list_lock);
832 list_add(&page->lru, &n->full);
833 spin_unlock(&n->list_lock);
836 static void remove_full(struct kmem_cache *s, struct page *page)
838 struct kmem_cache_node *n;
840 if (!(s->flags & SLAB_STORE_USER))
843 n = get_node(s, page_to_nid(page));
845 spin_lock(&n->list_lock);
846 list_del(&page->lru);
847 spin_unlock(&n->list_lock);
850 /* Tracking of the number of slabs for debugging purposes */
851 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
853 struct kmem_cache_node *n = get_node(s, node);
855 return atomic_long_read(&n->nr_slabs);
858 static inline void inc_slabs_node(struct kmem_cache *s, int node)
860 struct kmem_cache_node *n = get_node(s, node);
863 * May be called early in order to allocate a slab for the
864 * kmem_cache_node structure. Solve the chicken-egg
865 * dilemma by deferring the increment of the count during
866 * bootstrap (see early_kmem_cache_node_alloc).
868 if (!NUMA_BUILD || n)
869 atomic_long_inc(&n->nr_slabs);
871 static inline void dec_slabs_node(struct kmem_cache *s, int node)
873 struct kmem_cache_node *n = get_node(s, node);
875 atomic_long_dec(&n->nr_slabs);
878 /* Object debug checks for alloc/free paths */
879 static void setup_object_debug(struct kmem_cache *s, struct page *page,
882 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
885 init_object(s, object, 0);
886 init_tracking(s, object);
889 static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
890 void *object, void *addr)
892 if (!check_slab(s, page))
895 if (!on_freelist(s, page, object)) {
896 object_err(s, page, object, "Object already allocated");
900 if (!check_valid_pointer(s, page, object)) {
901 object_err(s, page, object, "Freelist Pointer check fails");
905 if (!check_object(s, page, object, 0))
908 /* Success perform special debug activities for allocs */
909 if (s->flags & SLAB_STORE_USER)
910 set_track(s, object, TRACK_ALLOC, addr);
911 trace(s, page, object, 1);
912 init_object(s, object, 1);
916 if (PageSlab(page)) {
918 * If this is a slab page then lets do the best we can
919 * to avoid issues in the future. Marking all objects
920 * as used avoids touching the remaining objects.
922 slab_fix(s, "Marking all objects used");
923 page->inuse = page->objects;
924 page->freelist = NULL;
929 static int free_debug_processing(struct kmem_cache *s, struct page *page,
930 void *object, void *addr)
932 if (!check_slab(s, page))
935 if (!check_valid_pointer(s, page, object)) {
936 slab_err(s, page, "Invalid object pointer 0x%p", object);
940 if (on_freelist(s, page, object)) {
941 object_err(s, page, object, "Object already free");
945 if (!check_object(s, page, object, 1))
948 if (unlikely(s != page->slab)) {
949 if (!PageSlab(page)) {
950 slab_err(s, page, "Attempt to free object(0x%p) "
951 "outside of slab", object);
952 } else if (!page->slab) {
954 "SLUB <none>: no slab for object 0x%p.\n",
958 object_err(s, page, object,
959 "page slab pointer corrupt.");
963 /* Special debug activities for freeing objects */
964 if (!SlabFrozen(page) && !page->freelist)
965 remove_full(s, page);
966 if (s->flags & SLAB_STORE_USER)
967 set_track(s, object, TRACK_FREE, addr);
968 trace(s, page, object, 0);
969 init_object(s, object, 0);
973 slab_fix(s, "Object at 0x%p not freed", object);
977 static int __init setup_slub_debug(char *str)
979 slub_debug = DEBUG_DEFAULT_FLAGS;
980 if (*str++ != '=' || !*str)
982 * No options specified. Switch on full debugging.
988 * No options but restriction on slabs. This means full
989 * debugging for slabs matching a pattern.
996 * Switch off all debugging measures.
1001 * Determine which debug features should be switched on
1003 for (; *str && *str != ','; str++) {
1004 switch (tolower(*str)) {
1006 slub_debug |= SLAB_DEBUG_FREE;
1009 slub_debug |= SLAB_RED_ZONE;
1012 slub_debug |= SLAB_POISON;
1015 slub_debug |= SLAB_STORE_USER;
1018 slub_debug |= SLAB_TRACE;
1021 printk(KERN_ERR "slub_debug option '%c' "
1022 "unknown. skipped\n", *str);
1028 slub_debug_slabs = str + 1;
1033 __setup("slub_debug", setup_slub_debug);
1035 static unsigned long kmem_cache_flags(unsigned long objsize,
1036 unsigned long flags, const char *name,
1037 void (*ctor)(struct kmem_cache *, void *))
1040 * Enable debugging if selected on the kernel commandline.
1042 if (slub_debug && (!slub_debug_slabs ||
1043 strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)) == 0))
1044 flags |= slub_debug;
1049 static inline void setup_object_debug(struct kmem_cache *s,
1050 struct page *page, void *object) {}
1052 static inline int alloc_debug_processing(struct kmem_cache *s,
1053 struct page *page, void *object, void *addr) { return 0; }
1055 static inline int free_debug_processing(struct kmem_cache *s,
1056 struct page *page, void *object, void *addr) { return 0; }
1058 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1060 static inline int check_object(struct kmem_cache *s, struct page *page,
1061 void *object, int active) { return 1; }
1062 static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
1063 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1064 unsigned long flags, const char *name,
1065 void (*ctor)(struct kmem_cache *, void *))
1069 #define slub_debug 0
1071 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1073 static inline void inc_slabs_node(struct kmem_cache *s, int node) {}
1074 static inline void dec_slabs_node(struct kmem_cache *s, int node) {}
1077 * Slab allocation and freeing
1079 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1082 int pages = 1 << s->order;
1084 flags |= s->allocflags;
1087 page = alloc_pages(flags, s->order);
1089 page = alloc_pages_node(node, flags, s->order);
1094 page->objects = s->objects;
1095 mod_zone_page_state(page_zone(page),
1096 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1097 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1103 static void setup_object(struct kmem_cache *s, struct page *page,
1106 setup_object_debug(s, page, object);
1107 if (unlikely(s->ctor))
1111 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1118 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1120 page = allocate_slab(s,
1121 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1125 inc_slabs_node(s, page_to_nid(page));
1127 page->flags |= 1 << PG_slab;
1128 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1129 SLAB_STORE_USER | SLAB_TRACE))
1132 start = page_address(page);
1134 if (unlikely(s->flags & SLAB_POISON))
1135 memset(start, POISON_INUSE, PAGE_SIZE << s->order);
1138 for_each_object(p, s, start) {
1139 setup_object(s, page, last);
1140 set_freepointer(s, last, p);
1143 setup_object(s, page, last);
1144 set_freepointer(s, last, NULL);
1146 page->freelist = start;
1152 static void __free_slab(struct kmem_cache *s, struct page *page)
1154 int pages = 1 << s->order;
1156 if (unlikely(SlabDebug(page))) {
1159 slab_pad_check(s, page);
1160 for_each_object(p, s, page_address(page))
1161 check_object(s, page, p, 0);
1162 ClearSlabDebug(page);
1165 mod_zone_page_state(page_zone(page),
1166 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1167 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1170 __ClearPageSlab(page);
1171 reset_page_mapcount(page);
1172 __free_pages(page, s->order);
1175 static void rcu_free_slab(struct rcu_head *h)
1179 page = container_of((struct list_head *)h, struct page, lru);
1180 __free_slab(page->slab, page);
1183 static void free_slab(struct kmem_cache *s, struct page *page)
1185 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1187 * RCU free overloads the RCU head over the LRU
1189 struct rcu_head *head = (void *)&page->lru;
1191 call_rcu(head, rcu_free_slab);
1193 __free_slab(s, page);
1196 static void discard_slab(struct kmem_cache *s, struct page *page)
1198 dec_slabs_node(s, page_to_nid(page));
1203 * Per slab locking using the pagelock
1205 static __always_inline void slab_lock(struct page *page)
1207 bit_spin_lock(PG_locked, &page->flags);
1210 static __always_inline void slab_unlock(struct page *page)
1212 __bit_spin_unlock(PG_locked, &page->flags);
1215 static __always_inline int slab_trylock(struct page *page)
1219 rc = bit_spin_trylock(PG_locked, &page->flags);
1224 * Management of partially allocated slabs
1226 static void add_partial(struct kmem_cache_node *n,
1227 struct page *page, int tail)
1229 spin_lock(&n->list_lock);
1232 list_add_tail(&page->lru, &n->partial);
1234 list_add(&page->lru, &n->partial);
1235 spin_unlock(&n->list_lock);
1238 static void remove_partial(struct kmem_cache *s,
1241 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1243 spin_lock(&n->list_lock);
1244 list_del(&page->lru);
1246 spin_unlock(&n->list_lock);
1250 * Lock slab and remove from the partial list.
1252 * Must hold list_lock.
1254 static inline int lock_and_freeze_slab(struct kmem_cache_node *n, struct page *page)
1256 if (slab_trylock(page)) {
1257 list_del(&page->lru);
1259 SetSlabFrozen(page);
1266 * Try to allocate a partial slab from a specific node.
1268 static struct page *get_partial_node(struct kmem_cache_node *n)
1273 * Racy check. If we mistakenly see no partial slabs then we
1274 * just allocate an empty slab. If we mistakenly try to get a
1275 * partial slab and there is none available then get_partials()
1278 if (!n || !n->nr_partial)
1281 spin_lock(&n->list_lock);
1282 list_for_each_entry(page, &n->partial, lru)
1283 if (lock_and_freeze_slab(n, page))
1287 spin_unlock(&n->list_lock);
1292 * Get a page from somewhere. Search in increasing NUMA distances.
1294 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1297 struct zonelist *zonelist;
1302 * The defrag ratio allows a configuration of the tradeoffs between
1303 * inter node defragmentation and node local allocations. A lower
1304 * defrag_ratio increases the tendency to do local allocations
1305 * instead of attempting to obtain partial slabs from other nodes.
1307 * If the defrag_ratio is set to 0 then kmalloc() always
1308 * returns node local objects. If the ratio is higher then kmalloc()
1309 * may return off node objects because partial slabs are obtained
1310 * from other nodes and filled up.
1312 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1313 * defrag_ratio = 1000) then every (well almost) allocation will
1314 * first attempt to defrag slab caches on other nodes. This means
1315 * scanning over all nodes to look for partial slabs which may be
1316 * expensive if we do it every time we are trying to find a slab
1317 * with available objects.
1319 if (!s->remote_node_defrag_ratio ||
1320 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1323 zonelist = &NODE_DATA(
1324 slab_node(current->mempolicy))->node_zonelists[gfp_zone(flags)];
1325 for (z = zonelist->zones; *z; z++) {
1326 struct kmem_cache_node *n;
1328 n = get_node(s, zone_to_nid(*z));
1330 if (n && cpuset_zone_allowed_hardwall(*z, flags) &&
1331 n->nr_partial > MIN_PARTIAL) {
1332 page = get_partial_node(n);
1342 * Get a partial page, lock it and return it.
1344 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1347 int searchnode = (node == -1) ? numa_node_id() : node;
1349 page = get_partial_node(get_node(s, searchnode));
1350 if (page || (flags & __GFP_THISNODE))
1353 return get_any_partial(s, flags);
1357 * Move a page back to the lists.
1359 * Must be called with the slab lock held.
1361 * On exit the slab lock will have been dropped.
1363 static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
1365 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1366 struct kmem_cache_cpu *c = get_cpu_slab(s, smp_processor_id());
1368 ClearSlabFrozen(page);
1371 if (page->freelist) {
1372 add_partial(n, page, tail);
1373 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1375 stat(c, DEACTIVATE_FULL);
1376 if (SlabDebug(page) && (s->flags & SLAB_STORE_USER))
1381 stat(c, DEACTIVATE_EMPTY);
1382 if (n->nr_partial < MIN_PARTIAL) {
1384 * Adding an empty slab to the partial slabs in order
1385 * to avoid page allocator overhead. This slab needs
1386 * to come after the other slabs with objects in
1387 * so that the others get filled first. That way the
1388 * size of the partial list stays small.
1390 * kmem_cache_shrink can reclaim any empty slabs from the
1393 add_partial(n, page, 1);
1397 stat(get_cpu_slab(s, raw_smp_processor_id()), FREE_SLAB);
1398 discard_slab(s, page);
1404 * Remove the cpu slab
1406 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1408 struct page *page = c->page;
1412 stat(c, DEACTIVATE_REMOTE_FREES);
1414 * Merge cpu freelist into slab freelist. Typically we get here
1415 * because both freelists are empty. So this is unlikely
1418 while (unlikely(c->freelist)) {
1421 tail = 0; /* Hot objects. Put the slab first */
1423 /* Retrieve object from cpu_freelist */
1424 object = c->freelist;
1425 c->freelist = c->freelist[c->offset];
1427 /* And put onto the regular freelist */
1428 object[c->offset] = page->freelist;
1429 page->freelist = object;
1433 unfreeze_slab(s, page, tail);
1436 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1438 stat(c, CPUSLAB_FLUSH);
1440 deactivate_slab(s, c);
1446 * Called from IPI handler with interrupts disabled.
1448 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1450 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1452 if (likely(c && c->page))
1456 static void flush_cpu_slab(void *d)
1458 struct kmem_cache *s = d;
1460 __flush_cpu_slab(s, smp_processor_id());
1463 static void flush_all(struct kmem_cache *s)
1466 on_each_cpu(flush_cpu_slab, s, 1, 1);
1468 unsigned long flags;
1470 local_irq_save(flags);
1472 local_irq_restore(flags);
1477 * Check if the objects in a per cpu structure fit numa
1478 * locality expectations.
1480 static inline int node_match(struct kmem_cache_cpu *c, int node)
1483 if (node != -1 && c->node != node)
1490 * Slow path. The lockless freelist is empty or we need to perform
1493 * Interrupts are disabled.
1495 * Processing is still very fast if new objects have been freed to the
1496 * regular freelist. In that case we simply take over the regular freelist
1497 * as the lockless freelist and zap the regular freelist.
1499 * If that is not working then we fall back to the partial lists. We take the
1500 * first element of the freelist as the object to allocate now and move the
1501 * rest of the freelist to the lockless freelist.
1503 * And if we were unable to get a new slab from the partial slab lists then
1504 * we need to allocate a new slab. This is the slowest path since it involves
1505 * a call to the page allocator and the setup of a new slab.
1507 static void *__slab_alloc(struct kmem_cache *s,
1508 gfp_t gfpflags, int node, void *addr, struct kmem_cache_cpu *c)
1513 /* We handle __GFP_ZERO in the caller */
1514 gfpflags &= ~__GFP_ZERO;
1520 if (unlikely(!node_match(c, node)))
1523 stat(c, ALLOC_REFILL);
1526 object = c->page->freelist;
1527 if (unlikely(!object))
1529 if (unlikely(SlabDebug(c->page)))
1532 c->freelist = object[c->offset];
1533 c->page->inuse = c->page->objects;
1534 c->page->freelist = NULL;
1535 c->node = page_to_nid(c->page);
1537 slab_unlock(c->page);
1538 stat(c, ALLOC_SLOWPATH);
1542 deactivate_slab(s, c);
1545 new = get_partial(s, gfpflags, node);
1548 stat(c, ALLOC_FROM_PARTIAL);
1552 if (gfpflags & __GFP_WAIT)
1555 new = new_slab(s, gfpflags, node);
1557 if (gfpflags & __GFP_WAIT)
1558 local_irq_disable();
1561 c = get_cpu_slab(s, smp_processor_id());
1562 stat(c, ALLOC_SLAB);
1572 * No memory available.
1574 * If the slab uses higher order allocs but the object is
1575 * smaller than a page size then we can fallback in emergencies
1576 * to the page allocator via kmalloc_large. The page allocator may
1577 * have failed to obtain a higher order page and we can try to
1578 * allocate a single page if the object fits into a single page.
1579 * That is only possible if certain conditions are met that are being
1580 * checked when a slab is created.
1582 if (!(gfpflags & __GFP_NORETRY) &&
1583 (s->flags & __PAGE_ALLOC_FALLBACK)) {
1584 if (gfpflags & __GFP_WAIT)
1586 object = kmalloc_large(s->objsize, gfpflags);
1587 if (gfpflags & __GFP_WAIT)
1588 local_irq_disable();
1593 if (!alloc_debug_processing(s, c->page, object, addr))
1597 c->page->freelist = object[c->offset];
1603 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1604 * have the fastpath folded into their functions. So no function call
1605 * overhead for requests that can be satisfied on the fastpath.
1607 * The fastpath works by first checking if the lockless freelist can be used.
1608 * If not then __slab_alloc is called for slow processing.
1610 * Otherwise we can simply pick the next object from the lockless free list.
1612 static __always_inline void *slab_alloc(struct kmem_cache *s,
1613 gfp_t gfpflags, int node, void *addr)
1616 struct kmem_cache_cpu *c;
1617 unsigned long flags;
1619 local_irq_save(flags);
1620 c = get_cpu_slab(s, smp_processor_id());
1621 if (unlikely(!c->freelist || !node_match(c, node)))
1623 object = __slab_alloc(s, gfpflags, node, addr, c);
1626 object = c->freelist;
1627 c->freelist = object[c->offset];
1628 stat(c, ALLOC_FASTPATH);
1630 local_irq_restore(flags);
1632 if (unlikely((gfpflags & __GFP_ZERO) && object))
1633 memset(object, 0, c->objsize);
1638 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1640 return slab_alloc(s, gfpflags, -1, __builtin_return_address(0));
1642 EXPORT_SYMBOL(kmem_cache_alloc);
1645 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1647 return slab_alloc(s, gfpflags, node, __builtin_return_address(0));
1649 EXPORT_SYMBOL(kmem_cache_alloc_node);
1653 * Slow patch handling. This may still be called frequently since objects
1654 * have a longer lifetime than the cpu slabs in most processing loads.
1656 * So we still attempt to reduce cache line usage. Just take the slab
1657 * lock and free the item. If there is no additional partial page
1658 * handling required then we can return immediately.
1660 static void __slab_free(struct kmem_cache *s, struct page *page,
1661 void *x, void *addr, unsigned int offset)
1664 void **object = (void *)x;
1665 struct kmem_cache_cpu *c;
1667 c = get_cpu_slab(s, raw_smp_processor_id());
1668 stat(c, FREE_SLOWPATH);
1671 if (unlikely(SlabDebug(page)))
1675 prior = object[offset] = page->freelist;
1676 page->freelist = object;
1679 if (unlikely(SlabFrozen(page))) {
1680 stat(c, FREE_FROZEN);
1684 if (unlikely(!page->inuse))
1688 * Objects left in the slab. If it was not on the partial list before
1691 if (unlikely(!prior)) {
1692 add_partial(get_node(s, page_to_nid(page)), page, 1);
1693 stat(c, FREE_ADD_PARTIAL);
1703 * Slab still on the partial list.
1705 remove_partial(s, page);
1706 stat(c, FREE_REMOVE_PARTIAL);
1710 discard_slab(s, page);
1714 if (!free_debug_processing(s, page, x, addr))
1720 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1721 * can perform fastpath freeing without additional function calls.
1723 * The fastpath is only possible if we are freeing to the current cpu slab
1724 * of this processor. This typically the case if we have just allocated
1727 * If fastpath is not possible then fall back to __slab_free where we deal
1728 * with all sorts of special processing.
1730 static __always_inline void slab_free(struct kmem_cache *s,
1731 struct page *page, void *x, void *addr)
1733 void **object = (void *)x;
1734 struct kmem_cache_cpu *c;
1735 unsigned long flags;
1737 local_irq_save(flags);
1738 c = get_cpu_slab(s, smp_processor_id());
1739 debug_check_no_locks_freed(object, c->objsize);
1740 if (likely(page == c->page && c->node >= 0)) {
1741 object[c->offset] = c->freelist;
1742 c->freelist = object;
1743 stat(c, FREE_FASTPATH);
1745 __slab_free(s, page, x, addr, c->offset);
1747 local_irq_restore(flags);
1750 void kmem_cache_free(struct kmem_cache *s, void *x)
1754 page = virt_to_head_page(x);
1756 slab_free(s, page, x, __builtin_return_address(0));
1758 EXPORT_SYMBOL(kmem_cache_free);
1760 /* Figure out on which slab object the object resides */
1761 static struct page *get_object_page(const void *x)
1763 struct page *page = virt_to_head_page(x);
1765 if (!PageSlab(page))
1772 * Object placement in a slab is made very easy because we always start at
1773 * offset 0. If we tune the size of the object to the alignment then we can
1774 * get the required alignment by putting one properly sized object after
1777 * Notice that the allocation order determines the sizes of the per cpu
1778 * caches. Each processor has always one slab available for allocations.
1779 * Increasing the allocation order reduces the number of times that slabs
1780 * must be moved on and off the partial lists and is therefore a factor in
1785 * Mininum / Maximum order of slab pages. This influences locking overhead
1786 * and slab fragmentation. A higher order reduces the number of partial slabs
1787 * and increases the number of allocations possible without having to
1788 * take the list_lock.
1790 static int slub_min_order;
1791 static int slub_max_order = DEFAULT_MAX_ORDER;
1792 static int slub_min_objects = DEFAULT_MIN_OBJECTS;
1795 * Merge control. If this is set then no merging of slab caches will occur.
1796 * (Could be removed. This was introduced to pacify the merge skeptics.)
1798 static int slub_nomerge;
1801 * Calculate the order of allocation given an slab object size.
1803 * The order of allocation has significant impact on performance and other
1804 * system components. Generally order 0 allocations should be preferred since
1805 * order 0 does not cause fragmentation in the page allocator. Larger objects
1806 * be problematic to put into order 0 slabs because there may be too much
1807 * unused space left. We go to a higher order if more than 1/8th of the slab
1810 * In order to reach satisfactory performance we must ensure that a minimum
1811 * number of objects is in one slab. Otherwise we may generate too much
1812 * activity on the partial lists which requires taking the list_lock. This is
1813 * less a concern for large slabs though which are rarely used.
1815 * slub_max_order specifies the order where we begin to stop considering the
1816 * number of objects in a slab as critical. If we reach slub_max_order then
1817 * we try to keep the page order as low as possible. So we accept more waste
1818 * of space in favor of a small page order.
1820 * Higher order allocations also allow the placement of more objects in a
1821 * slab and thereby reduce object handling overhead. If the user has
1822 * requested a higher mininum order then we start with that one instead of
1823 * the smallest order which will fit the object.
1825 static inline int slab_order(int size, int min_objects,
1826 int max_order, int fract_leftover)
1830 int min_order = slub_min_order;
1832 if ((PAGE_SIZE << min_order) / size > 65535)
1833 return get_order(size * 65535) - 1;
1835 for (order = max(min_order,
1836 fls(min_objects * size - 1) - PAGE_SHIFT);
1837 order <= max_order; order++) {
1839 unsigned long slab_size = PAGE_SIZE << order;
1841 if (slab_size < min_objects * size)
1844 rem = slab_size % size;
1846 if (rem <= slab_size / fract_leftover)
1854 static inline int calculate_order(int size)
1861 * Attempt to find best configuration for a slab. This
1862 * works by first attempting to generate a layout with
1863 * the best configuration and backing off gradually.
1865 * First we reduce the acceptable waste in a slab. Then
1866 * we reduce the minimum objects required in a slab.
1868 min_objects = slub_min_objects;
1869 while (min_objects > 1) {
1871 while (fraction >= 4) {
1872 order = slab_order(size, min_objects,
1873 slub_max_order, fraction);
1874 if (order <= slub_max_order)
1882 * We were unable to place multiple objects in a slab. Now
1883 * lets see if we can place a single object there.
1885 order = slab_order(size, 1, slub_max_order, 1);
1886 if (order <= slub_max_order)
1890 * Doh this slab cannot be placed using slub_max_order.
1892 order = slab_order(size, 1, MAX_ORDER, 1);
1893 if (order <= MAX_ORDER)
1899 * Figure out what the alignment of the objects will be.
1901 static unsigned long calculate_alignment(unsigned long flags,
1902 unsigned long align, unsigned long size)
1905 * If the user wants hardware cache aligned objects then follow that
1906 * suggestion if the object is sufficiently large.
1908 * The hardware cache alignment cannot override the specified
1909 * alignment though. If that is greater then use it.
1911 if (flags & SLAB_HWCACHE_ALIGN) {
1912 unsigned long ralign = cache_line_size();
1913 while (size <= ralign / 2)
1915 align = max(align, ralign);
1918 if (align < ARCH_SLAB_MINALIGN)
1919 align = ARCH_SLAB_MINALIGN;
1921 return ALIGN(align, sizeof(void *));
1924 static void init_kmem_cache_cpu(struct kmem_cache *s,
1925 struct kmem_cache_cpu *c)
1930 c->offset = s->offset / sizeof(void *);
1931 c->objsize = s->objsize;
1932 #ifdef CONFIG_SLUB_STATS
1933 memset(c->stat, 0, NR_SLUB_STAT_ITEMS * sizeof(unsigned));
1937 static void init_kmem_cache_node(struct kmem_cache_node *n)
1940 spin_lock_init(&n->list_lock);
1941 INIT_LIST_HEAD(&n->partial);
1942 #ifdef CONFIG_SLUB_DEBUG
1943 atomic_long_set(&n->nr_slabs, 0);
1944 INIT_LIST_HEAD(&n->full);
1950 * Per cpu array for per cpu structures.
1952 * The per cpu array places all kmem_cache_cpu structures from one processor
1953 * close together meaning that it becomes possible that multiple per cpu
1954 * structures are contained in one cacheline. This may be particularly
1955 * beneficial for the kmalloc caches.
1957 * A desktop system typically has around 60-80 slabs. With 100 here we are
1958 * likely able to get per cpu structures for all caches from the array defined
1959 * here. We must be able to cover all kmalloc caches during bootstrap.
1961 * If the per cpu array is exhausted then fall back to kmalloc
1962 * of individual cachelines. No sharing is possible then.
1964 #define NR_KMEM_CACHE_CPU 100
1966 static DEFINE_PER_CPU(struct kmem_cache_cpu,
1967 kmem_cache_cpu)[NR_KMEM_CACHE_CPU];
1969 static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
1970 static cpumask_t kmem_cach_cpu_free_init_once = CPU_MASK_NONE;
1972 static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s,
1973 int cpu, gfp_t flags)
1975 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu);
1978 per_cpu(kmem_cache_cpu_free, cpu) =
1979 (void *)c->freelist;
1981 /* Table overflow: So allocate ourselves */
1983 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()),
1984 flags, cpu_to_node(cpu));
1989 init_kmem_cache_cpu(s, c);
1993 static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu)
1995 if (c < per_cpu(kmem_cache_cpu, cpu) ||
1996 c > per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) {
2000 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu);
2001 per_cpu(kmem_cache_cpu_free, cpu) = c;
2004 static void free_kmem_cache_cpus(struct kmem_cache *s)
2008 for_each_online_cpu(cpu) {
2009 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2012 s->cpu_slab[cpu] = NULL;
2013 free_kmem_cache_cpu(c, cpu);
2018 static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2022 for_each_online_cpu(cpu) {
2023 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2028 c = alloc_kmem_cache_cpu(s, cpu, flags);
2030 free_kmem_cache_cpus(s);
2033 s->cpu_slab[cpu] = c;
2039 * Initialize the per cpu array.
2041 static void init_alloc_cpu_cpu(int cpu)
2045 if (cpu_isset(cpu, kmem_cach_cpu_free_init_once))
2048 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--)
2049 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu);
2051 cpu_set(cpu, kmem_cach_cpu_free_init_once);
2054 static void __init init_alloc_cpu(void)
2058 for_each_online_cpu(cpu)
2059 init_alloc_cpu_cpu(cpu);
2063 static inline void free_kmem_cache_cpus(struct kmem_cache *s) {}
2064 static inline void init_alloc_cpu(void) {}
2066 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2068 init_kmem_cache_cpu(s, &s->cpu_slab);
2075 * No kmalloc_node yet so do it by hand. We know that this is the first
2076 * slab on the node for this slabcache. There are no concurrent accesses
2079 * Note that this function only works on the kmalloc_node_cache
2080 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2081 * memory on a fresh node that has no slab structures yet.
2083 static struct kmem_cache_node *early_kmem_cache_node_alloc(gfp_t gfpflags,
2087 struct kmem_cache_node *n;
2088 unsigned long flags;
2090 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2092 page = new_slab(kmalloc_caches, gfpflags, node);
2095 if (page_to_nid(page) != node) {
2096 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2098 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2099 "in order to be able to continue\n");
2104 page->freelist = get_freepointer(kmalloc_caches, n);
2106 kmalloc_caches->node[node] = n;
2107 #ifdef CONFIG_SLUB_DEBUG
2108 init_object(kmalloc_caches, n, 1);
2109 init_tracking(kmalloc_caches, n);
2111 init_kmem_cache_node(n);
2112 inc_slabs_node(kmalloc_caches, node);
2115 * lockdep requires consistent irq usage for each lock
2116 * so even though there cannot be a race this early in
2117 * the boot sequence, we still disable irqs.
2119 local_irq_save(flags);
2120 add_partial(n, page, 0);
2121 local_irq_restore(flags);
2125 static void free_kmem_cache_nodes(struct kmem_cache *s)
2129 for_each_node_state(node, N_NORMAL_MEMORY) {
2130 struct kmem_cache_node *n = s->node[node];
2131 if (n && n != &s->local_node)
2132 kmem_cache_free(kmalloc_caches, n);
2133 s->node[node] = NULL;
2137 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2142 if (slab_state >= UP)
2143 local_node = page_to_nid(virt_to_page(s));
2147 for_each_node_state(node, N_NORMAL_MEMORY) {
2148 struct kmem_cache_node *n;
2150 if (local_node == node)
2153 if (slab_state == DOWN) {
2154 n = early_kmem_cache_node_alloc(gfpflags,
2158 n = kmem_cache_alloc_node(kmalloc_caches,
2162 free_kmem_cache_nodes(s);
2168 init_kmem_cache_node(n);
2173 static void free_kmem_cache_nodes(struct kmem_cache *s)
2177 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2179 init_kmem_cache_node(&s->local_node);
2185 * calculate_sizes() determines the order and the distribution of data within
2188 static int calculate_sizes(struct kmem_cache *s)
2190 unsigned long flags = s->flags;
2191 unsigned long size = s->objsize;
2192 unsigned long align = s->align;
2195 * Round up object size to the next word boundary. We can only
2196 * place the free pointer at word boundaries and this determines
2197 * the possible location of the free pointer.
2199 size = ALIGN(size, sizeof(void *));
2201 #ifdef CONFIG_SLUB_DEBUG
2203 * Determine if we can poison the object itself. If the user of
2204 * the slab may touch the object after free or before allocation
2205 * then we should never poison the object itself.
2207 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2209 s->flags |= __OBJECT_POISON;
2211 s->flags &= ~__OBJECT_POISON;
2215 * If we are Redzoning then check if there is some space between the
2216 * end of the object and the free pointer. If not then add an
2217 * additional word to have some bytes to store Redzone information.
2219 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2220 size += sizeof(void *);
2224 * With that we have determined the number of bytes in actual use
2225 * by the object. This is the potential offset to the free pointer.
2229 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2232 * Relocate free pointer after the object if it is not
2233 * permitted to overwrite the first word of the object on
2236 * This is the case if we do RCU, have a constructor or
2237 * destructor or are poisoning the objects.
2240 size += sizeof(void *);
2243 #ifdef CONFIG_SLUB_DEBUG
2244 if (flags & SLAB_STORE_USER)
2246 * Need to store information about allocs and frees after
2249 size += 2 * sizeof(struct track);
2251 if (flags & SLAB_RED_ZONE)
2253 * Add some empty padding so that we can catch
2254 * overwrites from earlier objects rather than let
2255 * tracking information or the free pointer be
2256 * corrupted if an user writes before the start
2259 size += sizeof(void *);
2263 * Determine the alignment based on various parameters that the
2264 * user specified and the dynamic determination of cache line size
2267 align = calculate_alignment(flags, align, s->objsize);
2270 * SLUB stores one object immediately after another beginning from
2271 * offset 0. In order to align the objects we have to simply size
2272 * each object to conform to the alignment.
2274 size = ALIGN(size, align);
2277 if ((flags & __KMALLOC_CACHE) &&
2278 PAGE_SIZE / size < slub_min_objects) {
2280 * Kmalloc cache that would not have enough objects in
2281 * an order 0 page. Kmalloc slabs can fallback to
2282 * page allocator order 0 allocs so take a reasonably large
2283 * order that will allows us a good number of objects.
2285 s->order = max(slub_max_order, PAGE_ALLOC_COSTLY_ORDER);
2286 s->flags |= __PAGE_ALLOC_FALLBACK;
2287 s->allocflags |= __GFP_NOWARN;
2289 s->order = calculate_order(size);
2296 s->allocflags |= __GFP_COMP;
2298 if (s->flags & SLAB_CACHE_DMA)
2299 s->allocflags |= SLUB_DMA;
2301 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2302 s->allocflags |= __GFP_RECLAIMABLE;
2305 * Determine the number of objects per slab
2307 s->objects = (PAGE_SIZE << s->order) / size;
2309 return !!s->objects;
2313 static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2314 const char *name, size_t size,
2315 size_t align, unsigned long flags,
2316 void (*ctor)(struct kmem_cache *, void *))
2318 memset(s, 0, kmem_size);
2323 s->flags = kmem_cache_flags(size, flags, name, ctor);
2325 if (!calculate_sizes(s))
2330 s->remote_node_defrag_ratio = 100;
2332 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2335 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
2337 free_kmem_cache_nodes(s);
2339 if (flags & SLAB_PANIC)
2340 panic("Cannot create slab %s size=%lu realsize=%u "
2341 "order=%u offset=%u flags=%lx\n",
2342 s->name, (unsigned long)size, s->size, s->order,
2348 * Check if a given pointer is valid
2350 int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2354 page = get_object_page(object);
2356 if (!page || s != page->slab)
2357 /* No slab or wrong slab */
2360 if (!check_valid_pointer(s, page, object))
2364 * We could also check if the object is on the slabs freelist.
2365 * But this would be too expensive and it seems that the main
2366 * purpose of kmem_ptr_valid() is to check if the object belongs
2367 * to a certain slab.
2371 EXPORT_SYMBOL(kmem_ptr_validate);
2374 * Determine the size of a slab object
2376 unsigned int kmem_cache_size(struct kmem_cache *s)
2380 EXPORT_SYMBOL(kmem_cache_size);
2382 const char *kmem_cache_name(struct kmem_cache *s)
2386 EXPORT_SYMBOL(kmem_cache_name);
2388 static void list_slab_objects(struct kmem_cache *s, struct page *page,
2391 #ifdef CONFIG_SLUB_DEBUG
2392 void *addr = page_address(page);
2394 DECLARE_BITMAP(map, page->objects);
2396 bitmap_zero(map, page->objects);
2397 slab_err(s, page, "%s", text);
2399 for_each_free_object(p, s, page->freelist)
2400 set_bit(slab_index(p, s, addr), map);
2402 for_each_object(p, s, addr, page->objects) {
2404 if (!test_bit(slab_index(p, s, addr), map)) {
2405 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2407 print_tracking(s, p);
2415 * Attempt to free all partial slabs on a node.
2417 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
2419 unsigned long flags;
2420 struct page *page, *h;
2422 spin_lock_irqsave(&n->list_lock, flags);
2423 list_for_each_entry_safe(page, h, &n->partial, lru) {
2425 list_del(&page->lru);
2426 discard_slab(s, page);
2429 list_slab_objects(s, page,
2430 "Objects remaining on kmem_cache_close()");
2433 spin_unlock_irqrestore(&n->list_lock, flags);
2437 * Release all resources used by a slab cache.
2439 static inline int kmem_cache_close(struct kmem_cache *s)
2445 /* Attempt to free all objects */
2446 free_kmem_cache_cpus(s);
2447 for_each_node_state(node, N_NORMAL_MEMORY) {
2448 struct kmem_cache_node *n = get_node(s, node);
2451 if (n->nr_partial || slabs_node(s, node))
2454 free_kmem_cache_nodes(s);
2459 * Close a cache and release the kmem_cache structure
2460 * (must be used for caches created using kmem_cache_create)
2462 void kmem_cache_destroy(struct kmem_cache *s)
2464 down_write(&slub_lock);
2468 up_write(&slub_lock);
2469 if (kmem_cache_close(s)) {
2470 printk(KERN_ERR "SLUB %s: %s called for cache that "
2471 "still has objects.\n", s->name, __func__);
2474 sysfs_slab_remove(s);
2476 up_write(&slub_lock);
2478 EXPORT_SYMBOL(kmem_cache_destroy);
2480 /********************************************************************
2482 *******************************************************************/
2484 struct kmem_cache kmalloc_caches[PAGE_SHIFT + 1] __cacheline_aligned;
2485 EXPORT_SYMBOL(kmalloc_caches);
2487 static int __init setup_slub_min_order(char *str)
2489 get_option(&str, &slub_min_order);
2494 __setup("slub_min_order=", setup_slub_min_order);
2496 static int __init setup_slub_max_order(char *str)
2498 get_option(&str, &slub_max_order);
2503 __setup("slub_max_order=", setup_slub_max_order);
2505 static int __init setup_slub_min_objects(char *str)
2507 get_option(&str, &slub_min_objects);
2512 __setup("slub_min_objects=", setup_slub_min_objects);
2514 static int __init setup_slub_nomerge(char *str)
2520 __setup("slub_nomerge", setup_slub_nomerge);
2522 static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2523 const char *name, int size, gfp_t gfp_flags)
2525 unsigned int flags = 0;
2527 if (gfp_flags & SLUB_DMA)
2528 flags = SLAB_CACHE_DMA;
2530 down_write(&slub_lock);
2531 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
2532 flags | __KMALLOC_CACHE, NULL))
2535 list_add(&s->list, &slab_caches);
2536 up_write(&slub_lock);
2537 if (sysfs_slab_add(s))
2542 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2545 #ifdef CONFIG_ZONE_DMA
2546 static struct kmem_cache *kmalloc_caches_dma[PAGE_SHIFT + 1];
2548 static void sysfs_add_func(struct work_struct *w)
2550 struct kmem_cache *s;
2552 down_write(&slub_lock);
2553 list_for_each_entry(s, &slab_caches, list) {
2554 if (s->flags & __SYSFS_ADD_DEFERRED) {
2555 s->flags &= ~__SYSFS_ADD_DEFERRED;
2559 up_write(&slub_lock);
2562 static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2564 static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2566 struct kmem_cache *s;
2570 s = kmalloc_caches_dma[index];
2574 /* Dynamically create dma cache */
2575 if (flags & __GFP_WAIT)
2576 down_write(&slub_lock);
2578 if (!down_write_trylock(&slub_lock))
2582 if (kmalloc_caches_dma[index])
2585 realsize = kmalloc_caches[index].objsize;
2586 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2587 (unsigned int)realsize);
2588 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2590 if (!s || !text || !kmem_cache_open(s, flags, text,
2591 realsize, ARCH_KMALLOC_MINALIGN,
2592 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) {
2598 list_add(&s->list, &slab_caches);
2599 kmalloc_caches_dma[index] = s;
2601 schedule_work(&sysfs_add_work);
2604 up_write(&slub_lock);
2606 return kmalloc_caches_dma[index];
2611 * Conversion table for small slabs sizes / 8 to the index in the
2612 * kmalloc array. This is necessary for slabs < 192 since we have non power
2613 * of two cache sizes there. The size of larger slabs can be determined using
2616 static s8 size_index[24] = {
2643 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2649 return ZERO_SIZE_PTR;
2651 index = size_index[(size - 1) / 8];
2653 index = fls(size - 1);
2655 #ifdef CONFIG_ZONE_DMA
2656 if (unlikely((flags & SLUB_DMA)))
2657 return dma_kmalloc_cache(index, flags);
2660 return &kmalloc_caches[index];
2663 void *__kmalloc(size_t size, gfp_t flags)
2665 struct kmem_cache *s;
2667 if (unlikely(size > PAGE_SIZE))
2668 return kmalloc_large(size, flags);
2670 s = get_slab(size, flags);
2672 if (unlikely(ZERO_OR_NULL_PTR(s)))
2675 return slab_alloc(s, flags, -1, __builtin_return_address(0));
2677 EXPORT_SYMBOL(__kmalloc);
2679 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2681 struct page *page = alloc_pages_node(node, flags | __GFP_COMP,
2685 return page_address(page);
2691 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2693 struct kmem_cache *s;
2695 if (unlikely(size > PAGE_SIZE))
2696 return kmalloc_large_node(size, flags, node);
2698 s = get_slab(size, flags);
2700 if (unlikely(ZERO_OR_NULL_PTR(s)))
2703 return slab_alloc(s, flags, node, __builtin_return_address(0));
2705 EXPORT_SYMBOL(__kmalloc_node);
2708 size_t ksize(const void *object)
2711 struct kmem_cache *s;
2713 if (unlikely(object == ZERO_SIZE_PTR))
2716 page = virt_to_head_page(object);
2718 if (unlikely(!PageSlab(page)))
2719 return PAGE_SIZE << compound_order(page);
2723 #ifdef CONFIG_SLUB_DEBUG
2725 * Debugging requires use of the padding between object
2726 * and whatever may come after it.
2728 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2733 * If we have the need to store the freelist pointer
2734 * back there or track user information then we can
2735 * only use the space before that information.
2737 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2740 * Else we can use all the padding etc for the allocation
2744 EXPORT_SYMBOL(ksize);
2746 void kfree(const void *x)
2749 void *object = (void *)x;
2751 if (unlikely(ZERO_OR_NULL_PTR(x)))
2754 page = virt_to_head_page(x);
2755 if (unlikely(!PageSlab(page))) {
2759 slab_free(page->slab, page, object, __builtin_return_address(0));
2761 EXPORT_SYMBOL(kfree);
2764 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2765 * the remaining slabs by the number of items in use. The slabs with the
2766 * most items in use come first. New allocations will then fill those up
2767 * and thus they can be removed from the partial lists.
2769 * The slabs with the least items are placed last. This results in them
2770 * being allocated from last increasing the chance that the last objects
2771 * are freed in them.
2773 int kmem_cache_shrink(struct kmem_cache *s)
2777 struct kmem_cache_node *n;
2780 struct list_head *slabs_by_inuse =
2781 kmalloc(sizeof(struct list_head) * s->objects, GFP_KERNEL);
2782 unsigned long flags;
2784 if (!slabs_by_inuse)
2788 for_each_node_state(node, N_NORMAL_MEMORY) {
2789 n = get_node(s, node);
2794 for (i = 0; i < s->objects; i++)
2795 INIT_LIST_HEAD(slabs_by_inuse + i);
2797 spin_lock_irqsave(&n->list_lock, flags);
2800 * Build lists indexed by the items in use in each slab.
2802 * Note that concurrent frees may occur while we hold the
2803 * list_lock. page->inuse here is the upper limit.
2805 list_for_each_entry_safe(page, t, &n->partial, lru) {
2806 if (!page->inuse && slab_trylock(page)) {
2808 * Must hold slab lock here because slab_free
2809 * may have freed the last object and be
2810 * waiting to release the slab.
2812 list_del(&page->lru);
2815 discard_slab(s, page);
2817 list_move(&page->lru,
2818 slabs_by_inuse + page->inuse);
2823 * Rebuild the partial list with the slabs filled up most
2824 * first and the least used slabs at the end.
2826 for (i = s->objects - 1; i >= 0; i--)
2827 list_splice(slabs_by_inuse + i, n->partial.prev);
2829 spin_unlock_irqrestore(&n->list_lock, flags);
2832 kfree(slabs_by_inuse);
2835 EXPORT_SYMBOL(kmem_cache_shrink);
2837 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2838 static int slab_mem_going_offline_callback(void *arg)
2840 struct kmem_cache *s;
2842 down_read(&slub_lock);
2843 list_for_each_entry(s, &slab_caches, list)
2844 kmem_cache_shrink(s);
2845 up_read(&slub_lock);
2850 static void slab_mem_offline_callback(void *arg)
2852 struct kmem_cache_node *n;
2853 struct kmem_cache *s;
2854 struct memory_notify *marg = arg;
2857 offline_node = marg->status_change_nid;
2860 * If the node still has available memory. we need kmem_cache_node
2863 if (offline_node < 0)
2866 down_read(&slub_lock);
2867 list_for_each_entry(s, &slab_caches, list) {
2868 n = get_node(s, offline_node);
2871 * if n->nr_slabs > 0, slabs still exist on the node
2872 * that is going down. We were unable to free them,
2873 * and offline_pages() function shoudn't call this
2874 * callback. So, we must fail.
2876 BUG_ON(slabs_node(s, offline_node));
2878 s->node[offline_node] = NULL;
2879 kmem_cache_free(kmalloc_caches, n);
2882 up_read(&slub_lock);
2885 static int slab_mem_going_online_callback(void *arg)
2887 struct kmem_cache_node *n;
2888 struct kmem_cache *s;
2889 struct memory_notify *marg = arg;
2890 int nid = marg->status_change_nid;
2894 * If the node's memory is already available, then kmem_cache_node is
2895 * already created. Nothing to do.
2901 * We are bringing a node online. No memory is availabe yet. We must
2902 * allocate a kmem_cache_node structure in order to bring the node
2905 down_read(&slub_lock);
2906 list_for_each_entry(s, &slab_caches, list) {
2908 * XXX: kmem_cache_alloc_node will fallback to other nodes
2909 * since memory is not yet available from the node that
2912 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
2917 init_kmem_cache_node(n);
2921 up_read(&slub_lock);
2925 static int slab_memory_callback(struct notifier_block *self,
2926 unsigned long action, void *arg)
2931 case MEM_GOING_ONLINE:
2932 ret = slab_mem_going_online_callback(arg);
2934 case MEM_GOING_OFFLINE:
2935 ret = slab_mem_going_offline_callback(arg);
2938 case MEM_CANCEL_ONLINE:
2939 slab_mem_offline_callback(arg);
2942 case MEM_CANCEL_OFFLINE:
2946 ret = notifier_from_errno(ret);
2950 #endif /* CONFIG_MEMORY_HOTPLUG */
2952 /********************************************************************
2953 * Basic setup of slabs
2954 *******************************************************************/
2956 void __init kmem_cache_init(void)
2965 * Must first have the slab cache available for the allocations of the
2966 * struct kmem_cache_node's. There is special bootstrap code in
2967 * kmem_cache_open for slab_state == DOWN.
2969 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2970 sizeof(struct kmem_cache_node), GFP_KERNEL);
2971 kmalloc_caches[0].refcount = -1;
2974 hotplug_memory_notifier(slab_memory_callback, 1);
2977 /* Able to allocate the per node structures */
2978 slab_state = PARTIAL;
2980 /* Caches that are not of the two-to-the-power-of size */
2981 if (KMALLOC_MIN_SIZE <= 64) {
2982 create_kmalloc_cache(&kmalloc_caches[1],
2983 "kmalloc-96", 96, GFP_KERNEL);
2986 if (KMALLOC_MIN_SIZE <= 128) {
2987 create_kmalloc_cache(&kmalloc_caches[2],
2988 "kmalloc-192", 192, GFP_KERNEL);
2992 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++) {
2993 create_kmalloc_cache(&kmalloc_caches[i],
2994 "kmalloc", 1 << i, GFP_KERNEL);
3000 * Patch up the size_index table if we have strange large alignment
3001 * requirements for the kmalloc array. This is only the case for
3002 * MIPS it seems. The standard arches will not generate any code here.
3004 * Largest permitted alignment is 256 bytes due to the way we
3005 * handle the index determination for the smaller caches.
3007 * Make sure that nothing crazy happens if someone starts tinkering
3008 * around with ARCH_KMALLOC_MINALIGN
3010 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3011 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3013 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
3014 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
3018 /* Provide the correct kmalloc names now that the caches are up */
3019 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++)
3020 kmalloc_caches[i]. name =
3021 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
3024 register_cpu_notifier(&slab_notifier);
3025 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
3026 nr_cpu_ids * sizeof(struct kmem_cache_cpu *);
3028 kmem_size = sizeof(struct kmem_cache);
3032 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3033 " CPUs=%d, Nodes=%d\n",
3034 caches, cache_line_size(),
3035 slub_min_order, slub_max_order, slub_min_objects,
3036 nr_cpu_ids, nr_node_ids);
3040 * Find a mergeable slab cache
3042 static int slab_unmergeable(struct kmem_cache *s)
3044 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3047 if ((s->flags & __PAGE_ALLOC_FALLBACK))
3054 * We may have set a slab to be unmergeable during bootstrap.
3056 if (s->refcount < 0)
3062 static struct kmem_cache *find_mergeable(size_t size,
3063 size_t align, unsigned long flags, const char *name,
3064 void (*ctor)(struct kmem_cache *, void *))
3066 struct kmem_cache *s;
3068 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3074 size = ALIGN(size, sizeof(void *));
3075 align = calculate_alignment(flags, align, size);
3076 size = ALIGN(size, align);
3077 flags = kmem_cache_flags(size, flags, name, NULL);
3079 list_for_each_entry(s, &slab_caches, list) {
3080 if (slab_unmergeable(s))
3086 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3089 * Check if alignment is compatible.
3090 * Courtesy of Adrian Drzewiecki
3092 if ((s->size & ~(align - 1)) != s->size)
3095 if (s->size - size >= sizeof(void *))
3103 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3104 size_t align, unsigned long flags,
3105 void (*ctor)(struct kmem_cache *, void *))
3107 struct kmem_cache *s;
3109 down_write(&slub_lock);
3110 s = find_mergeable(size, align, flags, name, ctor);
3116 * Adjust the object sizes so that we clear
3117 * the complete object on kzalloc.
3119 s->objsize = max(s->objsize, (int)size);
3122 * And then we need to update the object size in the
3123 * per cpu structures
3125 for_each_online_cpu(cpu)
3126 get_cpu_slab(s, cpu)->objsize = s->objsize;
3128 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3129 up_write(&slub_lock);
3131 if (sysfs_slab_alias(s, name))
3136 s = kmalloc(kmem_size, GFP_KERNEL);
3138 if (kmem_cache_open(s, GFP_KERNEL, name,
3139 size, align, flags, ctor)) {
3140 list_add(&s->list, &slab_caches);
3141 up_write(&slub_lock);
3142 if (sysfs_slab_add(s))
3148 up_write(&slub_lock);
3151 if (flags & SLAB_PANIC)
3152 panic("Cannot create slabcache %s\n", name);
3157 EXPORT_SYMBOL(kmem_cache_create);
3161 * Use the cpu notifier to insure that the cpu slabs are flushed when
3164 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3165 unsigned long action, void *hcpu)
3167 long cpu = (long)hcpu;
3168 struct kmem_cache *s;
3169 unsigned long flags;
3172 case CPU_UP_PREPARE:
3173 case CPU_UP_PREPARE_FROZEN:
3174 init_alloc_cpu_cpu(cpu);
3175 down_read(&slub_lock);
3176 list_for_each_entry(s, &slab_caches, list)
3177 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu,
3179 up_read(&slub_lock);
3182 case CPU_UP_CANCELED:
3183 case CPU_UP_CANCELED_FROZEN:
3185 case CPU_DEAD_FROZEN:
3186 down_read(&slub_lock);
3187 list_for_each_entry(s, &slab_caches, list) {
3188 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3190 local_irq_save(flags);
3191 __flush_cpu_slab(s, cpu);
3192 local_irq_restore(flags);
3193 free_kmem_cache_cpu(c, cpu);
3194 s->cpu_slab[cpu] = NULL;
3196 up_read(&slub_lock);
3204 static struct notifier_block __cpuinitdata slab_notifier = {
3205 .notifier_call = slab_cpuup_callback
3210 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller)
3212 struct kmem_cache *s;
3214 if (unlikely(size > PAGE_SIZE))
3215 return kmalloc_large(size, gfpflags);
3217 s = get_slab(size, gfpflags);
3219 if (unlikely(ZERO_OR_NULL_PTR(s)))
3222 return slab_alloc(s, gfpflags, -1, caller);
3225 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3226 int node, void *caller)
3228 struct kmem_cache *s;
3230 if (unlikely(size > PAGE_SIZE))
3231 return kmalloc_large_node(size, gfpflags, node);
3233 s = get_slab(size, gfpflags);
3235 if (unlikely(ZERO_OR_NULL_PTR(s)))
3238 return slab_alloc(s, gfpflags, node, caller);
3241 #if (defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)) || defined(CONFIG_SLABINFO)
3242 static unsigned long count_partial(struct kmem_cache_node *n)
3244 unsigned long flags;
3245 unsigned long x = 0;
3248 spin_lock_irqsave(&n->list_lock, flags);
3249 list_for_each_entry(page, &n->partial, lru)
3251 spin_unlock_irqrestore(&n->list_lock, flags);
3256 #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
3257 static int validate_slab(struct kmem_cache *s, struct page *page,
3261 void *addr = page_address(page);
3263 if (!check_slab(s, page) ||
3264 !on_freelist(s, page, NULL))
3267 /* Now we know that a valid freelist exists */
3268 bitmap_zero(map, page->objects);
3270 for_each_free_object(p, s, page->freelist) {
3271 set_bit(slab_index(p, s, addr), map);
3272 if (!check_object(s, page, p, 0))
3276 for_each_object(p, s, addr)
3277 if (!test_bit(slab_index(p, s, addr), map))
3278 if (!check_object(s, page, p, 1))
3283 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3286 if (slab_trylock(page)) {
3287 validate_slab(s, page, map);
3290 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3293 if (s->flags & DEBUG_DEFAULT_FLAGS) {
3294 if (!SlabDebug(page))
3295 printk(KERN_ERR "SLUB %s: SlabDebug not set "
3296 "on slab 0x%p\n", s->name, page);
3298 if (SlabDebug(page))
3299 printk(KERN_ERR "SLUB %s: SlabDebug set on "
3300 "slab 0x%p\n", s->name, page);
3304 static int validate_slab_node(struct kmem_cache *s,
3305 struct kmem_cache_node *n, unsigned long *map)
3307 unsigned long count = 0;
3309 unsigned long flags;
3311 spin_lock_irqsave(&n->list_lock, flags);
3313 list_for_each_entry(page, &n->partial, lru) {
3314 validate_slab_slab(s, page, map);
3317 if (count != n->nr_partial)
3318 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3319 "counter=%ld\n", s->name, count, n->nr_partial);
3321 if (!(s->flags & SLAB_STORE_USER))
3324 list_for_each_entry(page, &n->full, lru) {
3325 validate_slab_slab(s, page, map);
3328 if (count != atomic_long_read(&n->nr_slabs))
3329 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3330 "counter=%ld\n", s->name, count,
3331 atomic_long_read(&n->nr_slabs));
3334 spin_unlock_irqrestore(&n->list_lock, flags);
3338 static long validate_slab_cache(struct kmem_cache *s)
3341 unsigned long count = 0;
3342 unsigned long *map = kmalloc(BITS_TO_LONGS(s->objects) *
3343 sizeof(unsigned long), GFP_KERNEL);
3349 for_each_node_state(node, N_NORMAL_MEMORY) {
3350 struct kmem_cache_node *n = get_node(s, node);
3352 count += validate_slab_node(s, n, map);
3358 #ifdef SLUB_RESILIENCY_TEST
3359 static void resiliency_test(void)
3363 printk(KERN_ERR "SLUB resiliency testing\n");
3364 printk(KERN_ERR "-----------------------\n");
3365 printk(KERN_ERR "A. Corruption after allocation\n");
3367 p = kzalloc(16, GFP_KERNEL);
3369 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3370 " 0x12->0x%p\n\n", p + 16);
3372 validate_slab_cache(kmalloc_caches + 4);
3374 /* Hmmm... The next two are dangerous */
3375 p = kzalloc(32, GFP_KERNEL);
3376 p[32 + sizeof(void *)] = 0x34;
3377 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
3378 " 0x34 -> -0x%p\n", p);
3380 "If allocated object is overwritten then not detectable\n\n");
3382 validate_slab_cache(kmalloc_caches + 5);
3383 p = kzalloc(64, GFP_KERNEL);
3384 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3386 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3389 "If allocated object is overwritten then not detectable\n\n");
3390 validate_slab_cache(kmalloc_caches + 6);
3392 printk(KERN_ERR "\nB. Corruption after free\n");
3393 p = kzalloc(128, GFP_KERNEL);
3396 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3397 validate_slab_cache(kmalloc_caches + 7);
3399 p = kzalloc(256, GFP_KERNEL);
3402 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3404 validate_slab_cache(kmalloc_caches + 8);
3406 p = kzalloc(512, GFP_KERNEL);
3409 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3410 validate_slab_cache(kmalloc_caches + 9);
3413 static void resiliency_test(void) {};
3417 * Generate lists of code addresses where slabcache objects are allocated
3422 unsigned long count;
3435 unsigned long count;
3436 struct location *loc;
3439 static void free_loc_track(struct loc_track *t)
3442 free_pages((unsigned long)t->loc,
3443 get_order(sizeof(struct location) * t->max));
3446 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3451 order = get_order(sizeof(struct location) * max);
3453 l = (void *)__get_free_pages(flags, order);
3458 memcpy(l, t->loc, sizeof(struct location) * t->count);
3466 static int add_location(struct loc_track *t, struct kmem_cache *s,
3467 const struct track *track)
3469 long start, end, pos;
3472 unsigned long age = jiffies - track->when;
3478 pos = start + (end - start + 1) / 2;
3481 * There is nothing at "end". If we end up there
3482 * we need to add something to before end.
3487 caddr = t->loc[pos].addr;
3488 if (track->addr == caddr) {
3494 if (age < l->min_time)
3496 if (age > l->max_time)
3499 if (track->pid < l->min_pid)
3500 l->min_pid = track->pid;
3501 if (track->pid > l->max_pid)
3502 l->max_pid = track->pid;
3504 cpu_set(track->cpu, l->cpus);
3506 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3510 if (track->addr < caddr)
3517 * Not found. Insert new tracking element.
3519 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3525 (t->count - pos) * sizeof(struct location));
3528 l->addr = track->addr;
3532 l->min_pid = track->pid;
3533 l->max_pid = track->pid;
3534 cpus_clear(l->cpus);
3535 cpu_set(track->cpu, l->cpus);
3536 nodes_clear(l->nodes);
3537 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3541 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3542 struct page *page, enum track_item alloc)
3544 void *addr = page_address(page);
3545 DECLARE_BITMAP(map, page->objects);
3548 bitmap_zero(map, page->objects);
3549 for_each_free_object(p, s, page->freelist)
3550 set_bit(slab_index(p, s, addr), map);
3552 for_each_object(p, s, addr)
3553 if (!test_bit(slab_index(p, s, addr), map))
3554 add_location(t, s, get_track(s, p, alloc));
3557 static int list_locations(struct kmem_cache *s, char *buf,
3558 enum track_item alloc)
3562 struct loc_track t = { 0, 0, NULL };
3565 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3567 return sprintf(buf, "Out of memory\n");
3569 /* Push back cpu slabs */
3572 for_each_node_state(node, N_NORMAL_MEMORY) {
3573 struct kmem_cache_node *n = get_node(s, node);
3574 unsigned long flags;
3577 if (!atomic_long_read(&n->nr_slabs))
3580 spin_lock_irqsave(&n->list_lock, flags);
3581 list_for_each_entry(page, &n->partial, lru)
3582 process_slab(&t, s, page, alloc);
3583 list_for_each_entry(page, &n->full, lru)
3584 process_slab(&t, s, page, alloc);
3585 spin_unlock_irqrestore(&n->list_lock, flags);
3588 for (i = 0; i < t.count; i++) {
3589 struct location *l = &t.loc[i];
3591 if (len > PAGE_SIZE - 100)
3593 len += sprintf(buf + len, "%7ld ", l->count);
3596 len += sprint_symbol(buf + len, (unsigned long)l->addr);
3598 len += sprintf(buf + len, "<not-available>");
3600 if (l->sum_time != l->min_time) {
3601 unsigned long remainder;
3603 len += sprintf(buf + len, " age=%ld/%ld/%ld",
3605 div_long_long_rem(l->sum_time, l->count, &remainder),
3608 len += sprintf(buf + len, " age=%ld",
3611 if (l->min_pid != l->max_pid)
3612 len += sprintf(buf + len, " pid=%ld-%ld",
3613 l->min_pid, l->max_pid);
3615 len += sprintf(buf + len, " pid=%ld",
3618 if (num_online_cpus() > 1 && !cpus_empty(l->cpus) &&
3619 len < PAGE_SIZE - 60) {
3620 len += sprintf(buf + len, " cpus=");
3621 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3625 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
3626 len < PAGE_SIZE - 60) {
3627 len += sprintf(buf + len, " nodes=");
3628 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3632 len += sprintf(buf + len, "\n");
3637 len += sprintf(buf, "No data\n");
3641 enum slab_stat_type {
3648 #define SO_FULL (1 << SL_FULL)
3649 #define SO_PARTIAL (1 << SL_PARTIAL)
3650 #define SO_CPU (1 << SL_CPU)
3651 #define SO_OBJECTS (1 << SL_OBJECTS)
3653 static ssize_t show_slab_objects(struct kmem_cache *s,
3654 char *buf, unsigned long flags)
3656 unsigned long total = 0;
3660 unsigned long *nodes;
3661 unsigned long *per_cpu;
3663 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3666 per_cpu = nodes + nr_node_ids;
3668 for_each_possible_cpu(cpu) {
3670 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3680 if (flags & SO_CPU) {
3681 if (flags & SO_OBJECTS)
3692 for_each_node_state(node, N_NORMAL_MEMORY) {
3693 struct kmem_cache_node *n = get_node(s, node);
3695 if (flags & SO_PARTIAL) {
3696 if (flags & SO_OBJECTS)
3697 x = count_partial(n);
3704 if (flags & SO_FULL) {
3705 int full_slabs = atomic_long_read(&n->nr_slabs)
3709 if (flags & SO_OBJECTS)
3710 x = full_slabs * s->objects;
3718 x = sprintf(buf, "%lu", total);
3720 for_each_node_state(node, N_NORMAL_MEMORY)
3722 x += sprintf(buf + x, " N%d=%lu",
3726 return x + sprintf(buf + x, "\n");
3729 static int any_slab_objects(struct kmem_cache *s)
3734 for_each_possible_cpu(cpu) {
3735 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3741 for_each_online_node(node) {
3742 struct kmem_cache_node *n = get_node(s, node);
3747 if (n->nr_partial || atomic_long_read(&n->nr_slabs))
3753 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3754 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3756 struct slab_attribute {
3757 struct attribute attr;
3758 ssize_t (*show)(struct kmem_cache *s, char *buf);
3759 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3762 #define SLAB_ATTR_RO(_name) \
3763 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3765 #define SLAB_ATTR(_name) \
3766 static struct slab_attribute _name##_attr = \
3767 __ATTR(_name, 0644, _name##_show, _name##_store)
3769 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3771 return sprintf(buf, "%d\n", s->size);
3773 SLAB_ATTR_RO(slab_size);
3775 static ssize_t align_show(struct kmem_cache *s, char *buf)
3777 return sprintf(buf, "%d\n", s->align);
3779 SLAB_ATTR_RO(align);
3781 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3783 return sprintf(buf, "%d\n", s->objsize);
3785 SLAB_ATTR_RO(object_size);
3787 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3789 return sprintf(buf, "%d\n", s->objects);
3791 SLAB_ATTR_RO(objs_per_slab);
3793 static ssize_t order_show(struct kmem_cache *s, char *buf)
3795 return sprintf(buf, "%d\n", s->order);
3797 SLAB_ATTR_RO(order);
3799 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3802 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3804 return n + sprintf(buf + n, "\n");
3810 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3812 return sprintf(buf, "%d\n", s->refcount - 1);
3814 SLAB_ATTR_RO(aliases);
3816 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3818 return show_slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU);
3820 SLAB_ATTR_RO(slabs);
3822 static ssize_t partial_show(struct kmem_cache *s, char *buf)
3824 return show_slab_objects(s, buf, SO_PARTIAL);
3826 SLAB_ATTR_RO(partial);
3828 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3830 return show_slab_objects(s, buf, SO_CPU);
3832 SLAB_ATTR_RO(cpu_slabs);
3834 static ssize_t objects_show(struct kmem_cache *s, char *buf)
3836 return show_slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU|SO_OBJECTS);
3838 SLAB_ATTR_RO(objects);
3840 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3842 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3845 static ssize_t sanity_checks_store(struct kmem_cache *s,
3846 const char *buf, size_t length)
3848 s->flags &= ~SLAB_DEBUG_FREE;
3850 s->flags |= SLAB_DEBUG_FREE;
3853 SLAB_ATTR(sanity_checks);
3855 static ssize_t trace_show(struct kmem_cache *s, char *buf)
3857 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3860 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3863 s->flags &= ~SLAB_TRACE;
3865 s->flags |= SLAB_TRACE;
3870 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3872 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3875 static ssize_t reclaim_account_store(struct kmem_cache *s,
3876 const char *buf, size_t length)
3878 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3880 s->flags |= SLAB_RECLAIM_ACCOUNT;
3883 SLAB_ATTR(reclaim_account);
3885 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3887 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
3889 SLAB_ATTR_RO(hwcache_align);
3891 #ifdef CONFIG_ZONE_DMA
3892 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
3894 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
3896 SLAB_ATTR_RO(cache_dma);
3899 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
3901 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
3903 SLAB_ATTR_RO(destroy_by_rcu);
3905 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
3907 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
3910 static ssize_t red_zone_store(struct kmem_cache *s,
3911 const char *buf, size_t length)
3913 if (any_slab_objects(s))
3916 s->flags &= ~SLAB_RED_ZONE;
3918 s->flags |= SLAB_RED_ZONE;
3922 SLAB_ATTR(red_zone);
3924 static ssize_t poison_show(struct kmem_cache *s, char *buf)
3926 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
3929 static ssize_t poison_store(struct kmem_cache *s,
3930 const char *buf, size_t length)
3932 if (any_slab_objects(s))
3935 s->flags &= ~SLAB_POISON;
3937 s->flags |= SLAB_POISON;
3943 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
3945 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
3948 static ssize_t store_user_store(struct kmem_cache *s,
3949 const char *buf, size_t length)
3951 if (any_slab_objects(s))
3954 s->flags &= ~SLAB_STORE_USER;
3956 s->flags |= SLAB_STORE_USER;
3960 SLAB_ATTR(store_user);
3962 static ssize_t validate_show(struct kmem_cache *s, char *buf)
3967 static ssize_t validate_store(struct kmem_cache *s,
3968 const char *buf, size_t length)
3972 if (buf[0] == '1') {
3973 ret = validate_slab_cache(s);
3979 SLAB_ATTR(validate);
3981 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
3986 static ssize_t shrink_store(struct kmem_cache *s,
3987 const char *buf, size_t length)
3989 if (buf[0] == '1') {
3990 int rc = kmem_cache_shrink(s);
4000 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4002 if (!(s->flags & SLAB_STORE_USER))
4004 return list_locations(s, buf, TRACK_ALLOC);
4006 SLAB_ATTR_RO(alloc_calls);
4008 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4010 if (!(s->flags & SLAB_STORE_USER))
4012 return list_locations(s, buf, TRACK_FREE);
4014 SLAB_ATTR_RO(free_calls);
4017 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4019 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4022 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4023 const char *buf, size_t length)
4025 int n = simple_strtoul(buf, NULL, 10);
4028 s->remote_node_defrag_ratio = n * 10;
4031 SLAB_ATTR(remote_node_defrag_ratio);
4034 #ifdef CONFIG_SLUB_STATS
4035 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4037 unsigned long sum = 0;
4040 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4045 for_each_online_cpu(cpu) {
4046 unsigned x = get_cpu_slab(s, cpu)->stat[si];
4052 len = sprintf(buf, "%lu", sum);
4055 for_each_online_cpu(cpu) {
4056 if (data[cpu] && len < PAGE_SIZE - 20)
4057 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4061 return len + sprintf(buf + len, "\n");
4064 #define STAT_ATTR(si, text) \
4065 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4067 return show_stat(s, buf, si); \
4069 SLAB_ATTR_RO(text); \
4071 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4072 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4073 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4074 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4075 STAT_ATTR(FREE_FROZEN, free_frozen);
4076 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4077 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4078 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4079 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4080 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4081 STAT_ATTR(FREE_SLAB, free_slab);
4082 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4083 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4084 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4085 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4086 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4087 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4091 static struct attribute *slab_attrs[] = {
4092 &slab_size_attr.attr,
4093 &object_size_attr.attr,
4094 &objs_per_slab_attr.attr,
4099 &cpu_slabs_attr.attr,
4103 &sanity_checks_attr.attr,
4105 &hwcache_align_attr.attr,
4106 &reclaim_account_attr.attr,
4107 &destroy_by_rcu_attr.attr,
4108 &red_zone_attr.attr,
4110 &store_user_attr.attr,
4111 &validate_attr.attr,
4113 &alloc_calls_attr.attr,
4114 &free_calls_attr.attr,
4115 #ifdef CONFIG_ZONE_DMA
4116 &cache_dma_attr.attr,
4119 &remote_node_defrag_ratio_attr.attr,
4121 #ifdef CONFIG_SLUB_STATS
4122 &alloc_fastpath_attr.attr,
4123 &alloc_slowpath_attr.attr,
4124 &free_fastpath_attr.attr,
4125 &free_slowpath_attr.attr,
4126 &free_frozen_attr.attr,
4127 &free_add_partial_attr.attr,
4128 &free_remove_partial_attr.attr,
4129 &alloc_from_partial_attr.attr,
4130 &alloc_slab_attr.attr,
4131 &alloc_refill_attr.attr,
4132 &free_slab_attr.attr,
4133 &cpuslab_flush_attr.attr,
4134 &deactivate_full_attr.attr,
4135 &deactivate_empty_attr.attr,
4136 &deactivate_to_head_attr.attr,
4137 &deactivate_to_tail_attr.attr,
4138 &deactivate_remote_frees_attr.attr,
4143 static struct attribute_group slab_attr_group = {
4144 .attrs = slab_attrs,
4147 static ssize_t slab_attr_show(struct kobject *kobj,
4148 struct attribute *attr,
4151 struct slab_attribute *attribute;
4152 struct kmem_cache *s;
4155 attribute = to_slab_attr(attr);
4158 if (!attribute->show)
4161 err = attribute->show(s, buf);
4166 static ssize_t slab_attr_store(struct kobject *kobj,
4167 struct attribute *attr,
4168 const char *buf, size_t len)
4170 struct slab_attribute *attribute;
4171 struct kmem_cache *s;
4174 attribute = to_slab_attr(attr);
4177 if (!attribute->store)
4180 err = attribute->store(s, buf, len);
4185 static void kmem_cache_release(struct kobject *kobj)
4187 struct kmem_cache *s = to_slab(kobj);
4192 static struct sysfs_ops slab_sysfs_ops = {
4193 .show = slab_attr_show,
4194 .store = slab_attr_store,
4197 static struct kobj_type slab_ktype = {
4198 .sysfs_ops = &slab_sysfs_ops,
4199 .release = kmem_cache_release
4202 static int uevent_filter(struct kset *kset, struct kobject *kobj)
4204 struct kobj_type *ktype = get_ktype(kobj);
4206 if (ktype == &slab_ktype)
4211 static struct kset_uevent_ops slab_uevent_ops = {
4212 .filter = uevent_filter,
4215 static struct kset *slab_kset;
4217 #define ID_STR_LENGTH 64
4219 /* Create a unique string id for a slab cache:
4221 * Format :[flags-]size
4223 static char *create_unique_id(struct kmem_cache *s)
4225 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4232 * First flags affecting slabcache operations. We will only
4233 * get here for aliasable slabs so we do not need to support
4234 * too many flags. The flags here must cover all flags that
4235 * are matched during merging to guarantee that the id is
4238 if (s->flags & SLAB_CACHE_DMA)
4240 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4242 if (s->flags & SLAB_DEBUG_FREE)
4246 p += sprintf(p, "%07d", s->size);
4247 BUG_ON(p > name + ID_STR_LENGTH - 1);
4251 static int sysfs_slab_add(struct kmem_cache *s)
4257 if (slab_state < SYSFS)
4258 /* Defer until later */
4261 unmergeable = slab_unmergeable(s);
4264 * Slabcache can never be merged so we can use the name proper.
4265 * This is typically the case for debug situations. In that
4266 * case we can catch duplicate names easily.
4268 sysfs_remove_link(&slab_kset->kobj, s->name);
4272 * Create a unique name for the slab as a target
4275 name = create_unique_id(s);
4278 s->kobj.kset = slab_kset;
4279 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4281 kobject_put(&s->kobj);
4285 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4288 kobject_uevent(&s->kobj, KOBJ_ADD);
4290 /* Setup first alias */
4291 sysfs_slab_alias(s, s->name);
4297 static void sysfs_slab_remove(struct kmem_cache *s)
4299 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4300 kobject_del(&s->kobj);
4301 kobject_put(&s->kobj);
4305 * Need to buffer aliases during bootup until sysfs becomes
4306 * available lest we loose that information.
4308 struct saved_alias {
4309 struct kmem_cache *s;
4311 struct saved_alias *next;
4314 static struct saved_alias *alias_list;
4316 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4318 struct saved_alias *al;
4320 if (slab_state == SYSFS) {
4322 * If we have a leftover link then remove it.
4324 sysfs_remove_link(&slab_kset->kobj, name);
4325 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
4328 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4334 al->next = alias_list;
4339 static int __init slab_sysfs_init(void)
4341 struct kmem_cache *s;
4344 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
4346 printk(KERN_ERR "Cannot register slab subsystem.\n");
4352 list_for_each_entry(s, &slab_caches, list) {
4353 err = sysfs_slab_add(s);
4355 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4356 " to sysfs\n", s->name);
4359 while (alias_list) {
4360 struct saved_alias *al = alias_list;
4362 alias_list = alias_list->next;
4363 err = sysfs_slab_alias(al->s, al->name);
4365 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4366 " %s to sysfs\n", s->name);
4374 __initcall(slab_sysfs_init);
4378 * The /proc/slabinfo ABI
4380 #ifdef CONFIG_SLABINFO
4382 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4383 size_t count, loff_t *ppos)
4389 static void print_slabinfo_header(struct seq_file *m)
4391 seq_puts(m, "slabinfo - version: 2.1\n");
4392 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4393 "<objperslab> <pagesperslab>");
4394 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4395 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4399 static void *s_start(struct seq_file *m, loff_t *pos)
4403 down_read(&slub_lock);
4405 print_slabinfo_header(m);
4407 return seq_list_start(&slab_caches, *pos);
4410 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4412 return seq_list_next(p, &slab_caches, pos);
4415 static void s_stop(struct seq_file *m, void *p)
4417 up_read(&slub_lock);
4420 static int s_show(struct seq_file *m, void *p)
4422 unsigned long nr_partials = 0;
4423 unsigned long nr_slabs = 0;
4424 unsigned long nr_inuse = 0;
4425 unsigned long nr_objs;
4426 struct kmem_cache *s;
4429 s = list_entry(p, struct kmem_cache, list);
4431 for_each_online_node(node) {
4432 struct kmem_cache_node *n = get_node(s, node);
4437 nr_partials += n->nr_partial;
4438 nr_slabs += atomic_long_read(&n->nr_slabs);
4439 nr_inuse += count_partial(n);
4442 nr_objs = nr_slabs * s->objects;
4443 nr_inuse += (nr_slabs - nr_partials) * s->objects;
4445 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4446 nr_objs, s->size, s->objects, (1 << s->order));
4447 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4448 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4454 const struct seq_operations slabinfo_op = {
4461 #endif /* CONFIG_SLABINFO */